For months I've been hearing that Anthropic and OpenAI are doctrinal enemies, two opposite ways of building artificial intelligence, and from hearing it so often I almost believe it. One preaches that general AI will arrive if we accelerate. The other, that it'll arrive if we slow down just enough to align what we build. It sounds like a theological schism, like two churches that can't pray together. And then you look at the bills, the signed contracts, the servers each one rents, and the border between them turns more porous than the story admits. This spring Anthropic rented capacity on Elon Musk's supercomputer, the same Musk who three months earlier had publicly called it "evil". Meanwhile, Iranian drones were striking data centres in the Gulf. Neither of the two doctrinal stories cleanly predicts what the sector does when nobody's writing manifestos.
The 2021 split
Dario Amodei was VP of research at OpenAI. His sister Daniela held the VP of safety and policy. Tom Brown had led the team that built GPT-3, and alongside them left Sam McCandlish, Jared Kaplan, Chris Olah and other names from the safety and interpretability core. In late 2020 and early 2021 they leave the company as a bloc.
The public reason wasn't money. According to the versions circulating then, which their protagonists have refined in later interviews, the friction was fundamental: the suspicion that OpenAI was accelerating the deployment of capabilities — GPT-3 had just come out in public API — without investing proportionally in alignment and interpretability. They found Anthropic in early 2021, a Public Benefit Corporation from the start, with a declared mission of long-horizon safety research. And they begin where every doctrine that aspires to hold itself up begins: by drafting its founding texts.
Constitutional AI and the logic of the brake
The first of those texts is titled Constitutional AI, a paper from December 2022 proposing to align models without depending solely on direct human feedback (RLHF), the dominant method at OpenAI. The idea, stripped of its jargon: instead of humans voting on which of two answers is better case by case, the model is handed a "constitution" — an explicit set of carefully drafted normative principles — and trained to self-critique and correct itself against those principles.
What's interesting isn't that it works better, which is up for debate, but that it's more auditable. The constitution is a readable document. Anyone can read it, object to it, propose amendments; Anthropic publishes its own, grounded in the Universal Declaration of Human Rights, in AI safety principles and in no-harm guidelines. It isn't perfect and it has obvious gaps. But it's transparent in a way an aggregate of human votes never can be: there, there's nothing to read.
The second text is the Responsible Scaling Policy (RSP), first published in September 2023 and revised in successive versions since. The RSP is a voluntary self-regulation framework defining "AI Safety Levels" (ASL) inspired by the US biosafety levels, the ones that classify the handling of pathogens by their danger. Each level assumes a hypothesis about the model's capabilities — from harmless to potentially catastrophic — and demands containment measures to match: mandatory adversarial evaluation, access controls on the model's weights, incident-response protocols.
Beneath all that there's a simple bet. If the technical frontier approaches systems capable of dangerous things — agents with CBRN, cyber-offensive or self-replication skills — the protocols have to exist before and not after, because a certain class of harm admits no reverse. Alignment, in this logic, isn't a patch applied at the end but a prerequisite to the capability itself.
Learn by deploying, the opposite bet
Sam Altman has defended in interviews and on OpenAI's corporate blog the opposite thesis, and it's worth reading it without caricaturing it. He holds that the best way to make AI safe is to release it into the real world, watch how it behaves, correct in production and scale; that the protocols imagined in the lab are inferior to those refined by mass use; that waiting for formal guarantees before deploying means delaying the social benefit and handing the ground to less scrupulous competitors.
The stance has its logic, and it genuinely does. The deployment of ChatGPT in November 2022 generated an enormous amount of real feedback that improved later versions, and rapid iteration in production is a legitimate methodology in software, not an excuse. The problem, for its critics, is that some classes of harm in AI aren't reversed once they're loose. Iteration works when the harm is bounded and observable; it stumbles when it's diffuse, massive or irreversible. There the feedback arrives late.
The company's strategy has been consistent with its thesis: aggressive deployment, a widely accessible API, deep integration into Microsoft's products, successive generations of models pushing the capability frontier. The investment in safety exists, but deployment sets the pulse. And the user who opens Claude or ChatGPT every morning, without knowing it, is choosing one of the two liturgies. Which is better isn't settled by any benchmark; it's settled by which incentives you'd rather see win the next decade.
The alliances that don't fit the schism
Here the separation between the two begins to crumble, because both belong, financially, to an overlapping system of patronage that ties them more tightly than their doctrines let on.
OpenAI has Microsoft. The 2025 restructuring gave Microsoft a stake valued at around 135 billion dollars in the new OpenAI Group PBC, on a total committed investment of around 13 billion. The dependence has also been infrastructural: Azure was OpenAI's exclusive compute provider from 2019, an exclusivity lifted precisely with the October 2025 agreement, when Microsoft lost its right of first refusal over compute capacity. During those years, if Microsoft pulled the infrastructure, OpenAI was left without servers.
Anthropic has Amazon and Google, and here the split is surprising. Amazon has committed around 8 billion dollars across several rounds, the most-cited figure. But Google has put in considerably more: by 2026 its investments totalled at least 13 billion, with later commitments scaling well beyond that figure. Anthropic leans on AWS as its main infrastructure and complements it with Google Cloud's TPUs. The two hyperscalers have bet on both hypotheses at once. Microsoft, on speed. Amazon and Google, on caution. If one of the doctrines wins, the winner will be in the orbit of one of them; if both survive, all three get paid without choosing a side.
Looked at from above, the sector isn't a fight between Anthropic and OpenAI. It's a distribution of portfolios among three compute giants who learned long ago not to bet everything on one horse. The philosophical separation is real for the employees and for the users. For the underlying capital that moves the chips, it's diversification.
Musk's supercomputer
On 6 May 2026 something happened worth looking at slowly, because it belies a good part of the doctrinal story on both sides. Anthropic announced a deal with SpaceX to rent capacity on Colossus 1, xAI's data centre in Memphis and one of the largest supercomputers in the world. It will pay xAI 1.25 billion dollars a month until May 2029, which can total more than 40 billion, in exchange for access to more than 300 megawatts of compute. The deal includes, on another layer, an interest in exploring with SpaceX the development of compute capacity in orbit.
Three months earlier, Musk had written on X that Anthropic was "evil", "misanthropic" and that the lab "hated Western civilisation". He also has an active lawsuit against OpenAI for an alleged breach of its founding mission, and competes head-on with both through his own model, Grok. That he signs a 40-billion, three-year contract with Anthropic, and then publishes that he's spent time with the team and that "nobody set off his evil detector", says uncomfortable things about how the sector actually works.
The first is that the urgency for compute rules above any doctrinal consideration. A growth in usage that outstrips forecasts demands a capacity Amazon and Google alone don't deliver at the needed pace, and when the company is forced to look outside, the only provider able to serve 300 MW of frontier compute on schedule is xAI, whoever owns it and whatever it says in public. Silicon has no ideology, and neither does haste.
The second, more bitter, is that both the Responsible Scaling doctrine and Musk's warnings against Anthropic are, in part, stagecraft. When the moment comes to sign multi-billion contracts, the two parties sit at the same table. The public hostility is story; the contract is business. I don't say this out of cheap cynicism, but because that's how the sector operates when you take away the props.
And there's a third, which points to the structure of compute itself. If Anthropic now depends on four providers — AWS, Google Cloud, xAI/SpaceX and, in prospect, orbital capacity from SpaceX — its infrastructural concentration looks quite a lot like OpenAI's with Azure. The two preach differently and live on top of the same handful of physical providers. The doctrinal distinction blurs when the bottleneck is silicon and the silicon is shared out among three or four global companies.
The bombardments of the Persian Gulf
In February 2026, the United States and Israel launched a military operation against Iran. On 1 March, before dawn, Iranian Shahed drones struck two Amazon Web Services data centres in the United Arab Emirates; a third commercial centre in Bahrain was also hit, without it being clear whether it was a deliberate target or collateral damage. There was a second attack on an AWS centre in Bahrain on 1 April, and the next day Iranian state media claimed an attack on an Oracle centre in Dubai.
It's the first time in history that a state has deliberately bombed commercial data centres in an armed conflict. The Iranian regime had issued, on 31 March, a statement declaring the big American tech companies legitimate military targets: Microsoft, Google, Apple, Meta, Oracle, Intel, HP, IBM, Cisco, Dell, Palantir, Nvidia. The line separating civilian infrastructure from military target, for the AI sector, has just shifted.
This connects with the rivalry between Anthropic and OpenAI for a reason that appears in no manifesto. OpenAI's Stargate data centre in Abu Dhabi — the Emirati component, with G42 as local partner, of the global 500-billion-dollar project the company announced in 2025 alongside SoftBank, Oracle and MGX — is one of the assets exposed if the conflict drags on. The physical infrastructure of generative AI, sold as neutral and delocalised, has turned out to have geography, a border and hostile neighbours. When OpenAI and Anthropic compete for compute, they compete for specific buildings a state actor can bomb.
None of this figures in the listicles comparing Claude and ChatGPT, and yet it's a structural part of what happens when frontier compute concentrates in four or five locations on the planet and this century's war includes infrastructure among its targets. The two companies argue about how to build AI. Neither yet argues, with the clarity the matter demands, about where to build it.
The political labels and the pacts that belie them
A good part of the public discourse associates Anthropic with a progressive alignment, attentive to social biases and equity problems, and OpenAI with a more utilitarian one, technically neutral, focused on "mass deployment for aggregate benefit". The caricature has something true in it and a lot of simplification.
The true part: Claude's explicit constitutions pay attention to vulnerable groups, to demographic biases, to harm in sensitive contexts, and Anthropic's public communication underlines those points more than its rival's. The simplification comes when the contracts appear. When Anthropic rejected in February 2026 certain clauses of an agreement with the Department of Defense — concerning the use of Claude in mass surveillance and autonomous weapons — OpenAI signed with the Pentagon the next day, with different wording and barely hours after Altman said he shared Anthropic's stance. The two play on the same board with slightly different cards, and neither gets up from the table.
The political attributions, at the AI frontier, are marketing currency before they're a description of the practice. The user who chooses Claude out of ideological conviction without understanding the company's structural dependencies is buying, in part, a story the company itself can't sustain whole. The one who chooses ChatGPT convinced they're opting for the "pragmatic" thing doesn't fully see what they're accepting either. In their substrate, the two options resemble each other far more than their two explicit discourses admit.
What you'd rather see win
The rivalry between Anthropic and OpenAI, in what really affects the user, isn't settled by which model answers a riddle better. It's settled in the architecture of incentives. Whether you'd rather the sector advance under the logic of aggressive deployment, where production feedback governs and harms are corrected on the fly. Or under the logic of calibrated caution, where protocols precede deployment and theoretical harm weighs as much as observed harm. The two logics have virtues and bills: the first delivers more immediate utility, more speed and more short-term benefit, in exchange for more risk of harms nobody sees until it's late; the second delivers less speed and some missed opportunity, with a more expensive compliance operation, in exchange for less exposure to the unforeseen catastrophe.
Your choice of which tab to open each morning decides nothing on its own. But millions of aggregated choices — in subscriptions, in enterprise contracts, in APIs embedded inside workflows — do move market share, and market share, in a sector with margins choked by the cost of compute, decides which company can sustain its doctrine and which will have to betray it to keep from dying. You're a voter without knowing it every time you open a conversation. Meanwhile, outside the temple, the hyperscalers get paid and the drones keep flying.
Definiciones
Constitutional AI is Anthropic's framework for aligning language models through an explicit constitution — a drafted set of normative principles — instead of depending solely on aggregate human feedback. It was published in December 2022.
Responsible Scaling Policy (RSP) is Anthropic's voluntary policy defining AI safety levels (ASL) by analogy with biosafety levels. It was first published in September 2023 and has been revised in later versions.
Hyperscaler is a company with the capacity to operate data centres at a multi-billion global scale. In 2026 the main ones are Microsoft Azure, Amazon Web Services and Google Cloud, with xAI/SpaceX emerging as a fourth relevant player after the deal with Anthropic.
Stargate is the AI infrastructure project announced in 2025 by OpenAI alongside SoftBank, Oracle and MGX, with a declared investment of 500 billion dollars; its Emirati component in Abu Dhabi has G42 as local partner.
Referencias
Constitutional AI: Harmlessness from AI Feedback, by Bai et al. (Anthropic, 2022), is the paper proposing the constitution-based alignment method discussed in the article.
Anthropic's Responsible Scaling Policy (2023 and later revisions) is the source on the ASL levels: https://www.anthropic.com/news/anthropics-responsible-scaling-policy
Claude's Constitution (Anthropic) contains the public normative text of Claude's constitution: https://www.anthropic.com/constitution
The Wikipedia biographies of Dario Amodei and Daniela Amodei support the positions both held at OpenAI before founding Anthropic: https://en.wikipedia.org/wiki/Dario_Amodei and https://en.wikipedia.org/wiki/Daniela_Amodei
The Fortune report on OpenAI's restructuring (October 2025) is the basis for the valuation of Microsoft's stake: https://fortune.com/2025/10/28/openai-for-profit-restructuring-microsoft-stake/
The Axios analysis of Google's and Amazon's investments in Anthropic (April 2026) supports the figures for Anthropic's patronage: https://www.axios.com/2026/04/24/google-amazon-anthropic-investment
The Fortune report (May 2026) on Musk's turn regarding Anthropic accompanies the account of the compute deal: https://fortune.com/2026/05/07/spacex-anthropic-deal-elon-musk-ai-landlord-evil/
The CNBC note (May 2026) on the Anthropic–SpaceX deal, including the in-space development component, documents the terms: https://www.cnbc.com/2026/05/06/anthropic-spacex-data-center-capacity.html
The TechCrunch article (May 2026) details Anthropic's monthly payment to xAI for compute: https://techcrunch.com/2026/05/20/anthropic-will-pay-xai-1-25-billion-per-month-for-compute/
The Tech Policy Press / Just Security analysis (2026) examines the legal implications of the Iranian attacks on data centres in the UAE and Bahrain: https://www.techpolicy.press/the-legal-and-policy-fallout-from-data-center-strikes-in-the-middle-east-war/
The Fortune report (March 2026) on the Iranian drone attacks on Amazon data centres in the Gulf provides the chronology of the strikes: https://fortune.com/2026/03/09/irans-attacks-on-amazon-data-centers-in-uae-bahrain-signal-a-new-kind-of-war-as-ai-plays-an-increasingly-strategic-role-analysts-say/
The CNN and Bloomberg coverage (26 February 2026) of Anthropic's rejection of Department of Defense clauses grounds the military-contracts episode: https://www.cnn.com/2026/02/26/tech/anthropic-rejects-pentagon-offer
The official OpenAI page on the Stargate project and its Wikipedia entry detail the JV partners and the Emirati component: https://openai.com/index/announcing-the-stargate-project/ and https://en.wikipedia.org/wiki/Stargate_LLC
Russell, S. (2019), Human Compatible (Viking), and Christian, B. (2020), The Alignment Problem (Norton), are the two background references on the alignment problem.
También te interesa
- Who OpenAI really is
- DeepSeek and the end of the American monopoly
- Gemini, Claude and ChatGPT. Are you sure you have to choose?
No comments yet
No comments yet. Be the first.
Leave a comment